AI Health Answers and the Missing Dose

An inhalation reference value is the concentration at which a regulator has decided a substance in air starts to matter. It is the number behind every confident sentence about what is safe to breathe. For a great many of the compounds people actually sleep next to, it has never been published.

Our Exposure Ledger joins the compound Atlas to ten public exposure datasets. One of them is the Regulatory EPA Regional Screening Level table, which consolidates health-based values from IRIS, PPRTV, ATSDR, Cal EPA and OPP. Run the join and 78 compounds come back that are measured in the bodies of the general population and carry no inhalation reference value at all.

Of those 78 compounds, 48 are named by at least one chemical regulator, and 8 are prohibited outright in the European Union under the POPs Regulation. For none of them has anyone published a number describing how much in the air you breathe would matter. Inferred

That count is our own arithmetic on published federal and EU data, not a figure from a paper. It is reproducible from two files this site already serves: dose-gap.json lists the 78, and ledger.json carries each compound's regulatory record.

The easy version of this story is wrong

The easy version is that AI assistants get health facts wrong. The research does not really support it, and the part it does support is narrower and more unpleasant.

Andrikyan and colleagues queried an AI-powered search chatbot on ten common patient questions about the 50 most prescribed drugs in the United States, generating 500 answers. Median completeness and accuracy both came out at 100%. Peer-reviewed That is the finding people skip, and it is the one that matters most: these systems are usually right, which is exactly what makes the rest hard to catch.

On a preselected subset of 20 inaccurate answers, healthcare professionals judged 66% potentially harmful, and 22% capable of causing severe harm or even death if a patient followed the advice. Peer-reviewed A system that is nearly always right and occasionally dangerous is harder to use safely than one that is obviously unreliable, because nothing in the writing tells you which answer you are holding.

The wider literature is not yet in a state to settle this. Huo and colleagues systematically reviewed 137 studies evaluating chatbot health advice and found the reporting quality heterogeneous: 65% defined successful performance by subjective means, and fewer than a third addressed the ethical, regulatory and patient-safety implications of clinical integration. Peer-reviewed Their review exists to argue for a reporting standard, which is a fair description of where the field is.

Why an absence is the hardest answer to give

Here is the specific failure this site can speak to. A question shaped like how much X is safe has an expected shape of answer: a number, a unit, a comparison to a limit. Producing that shape is what these systems are built to do well. Establishing that no such number has ever been published is a different task, and nothing in the format rewards it.

So the honest answer for TDCPP is that no inhalation reference value exists, that it is on California's Proposition 65 list, and that these two facts are both true at once. That answer is unsatisfying, hard to phrase, and correct. Inferred We label this reading inferred rather than measured: it is how we read the behaviour of these systems, not something we have tested experimentally.

Formaldehyde is the contrast that makes the point. It has a reference value, decades of study, and a calibrated verdict on this site that says the exposure is real and worst when a product is new. A confident answer about formaldehyde is appropriate. The same confidence about the compound in the next paragraph of the same page is not, and nothing in the prose marks the transition.

The weakness in our own count

Our 78 depends entirely on treating the EPA table as the definition of “a published inhalation reference value.” It is the most consolidated source available, and it is not the world. Other jurisdictions publish values EPA has not adopted, and a compound absent from that table may have a number somewhere we did not look. The honest statement is narrower than the headline: no inhalation reference value in the source we checked, which is the most comprehensive single source we know of.

There is a second limit worth naming before anyone else does. We can say what has no published dose. We can say much less about what removes it: our ledger carries a sourced intervention for only a minority of these compounds, and that number describes the state of our own curation as much as the state of the evidence. We killed a headline stat once for exactly this reason, and we are not going to quietly reintroduce it here.

What this does not establish

A missing reference value is not evidence that a compound is dangerous. It is not evidence that it is safe either. It is evidence that the work has not been done, and treating that absence as reassurance and treating it as alarm are the same error pointed in opposite directions.

This distinction is the one that sank the best-known consumer chemical ranking of the last twenty years. Winter and Katz found that the methodology used to rank produce by pesticide contamination “lacks scientific credibility,” because it ranked presence while ignoring dose. Peer-reviewed That critique applies to us with equal force, which is why The Dose Gap ranks the missing dose-response itself rather than the compounds. Ranking ignorance is defensible precisely because ignorance is what it claims to rank.

Nor is this an argument that AI assistants are useless for this subject. They are good at the parts that are settled. The claim is narrower: they are least reliable exactly where this site's subject matter lives, because so much of it is unsettled.

What I would do

When an answer about a chemical in your home arrives with no source attached, the useful question is not whether it sounds authoritative. It is whether a number exists at all. Ask where the limit comes from. If the answer cannot name the body that set it, treat the confidence as a property of the writing rather than of the evidence.

That is the whole reason every claim on this site carries an evidence tier, and why the compound pages open with a verdict that is allowed to say depends. Most of ours do. Of the compounds carrying a verdict in the Atlas, the commonest stance is not that the risk is real; it is that it depends on conditions we then have to spell out. An answer that cannot express that gradient will always sound better than one that can.